MICCAI 2026: image quality assessment — the mathematical constructs

Why this page exists

These are the 11 MICCAI 2026 posters and orals tagged image quality assessment in MICCAI-2026-IQA/miccai2026-radiology-papers-2.csv. As on the domain adaptation page, each gets a short reading of its core idea and the mathematical objects it is built from: the score, loss or estimator, what it assumes, and which known tool it is an instance of. Results are left out except where a number makes a point. Each entry ends with What to watch, which names the assumption the maths leans on or the step that is heuristic rather than derived.

The tag is broad. Six papers grade or localise image quality directly (T-PM-174, W-AM-041, W-PM-008, W-PM-121, W-PM-139, W-PM-212). T-AM-063 grades annotation quality, T-AM-171 uses a quality classifier to steer a generator, W-AM-042 filters the quality of pseudo-labels, W-PM-123 identifies DICOM series as a quality-control step, and W-PM-194 has quality regression as one of four downstream tasks. Each entry opens by saying what is being graded.

Every entry has a widget on the interactive page. The widgets run on small toy numbers in the browser, not on the papers’ data, so they show mechanisms, not results. Where a toy exposes a gap in a paper, the entry says so.

Equations were reconstructed from pdftotext output. Where a paper gives a construct only in words, or its printed equation reads oddly, the entry says so rather than guessing silently.

What is being graded, and from what

The common thread is that quality labels are scarce, coarse or subjective, so most of the mathematics is about getting more out of weak labels: prototypes built from a handful of examples, positive-unlabelled risks, multiple-instance pooling, statistical filtering of noisy labels, and generators that manufacture paired data.

Construct map

A quick index by the kind of mathematics used. Several papers appear more than once.

Interactive figures

The interactive page has one widget per paper. Each has a callout naming the one thing worth trying first.

The papers

T-AM-063 · Decoupled Single-Mask Annotation Noise Detection via Cross-Sectional Patch Self-Consistency

Yinheng Zhu, Xiaowei Xu — arXiv:2607.05965 · MICCAI page · oral

What is graded. The annotation, not the image: which parts of a single vessel mask are unreliable.

Core idea. Coronary CT datasets carry one mask per scan, so there is no second rater to disagree with. The paper builds the second rater from the dataset itself. Vessels look alike in cross-section, so a 24 × 24 cross-sectional patch usually has near-identical twins elsewhere, in the same scan or in other patients. If two patches look the same but their masks differ, at least one mask is wrong. A per-bin z-score measures “differ more than usual”, and a median over all twins decides which patch is to blame. No network is trained to find the noise, so every flag comes with its evidence: the twin patches themselves.

Mathematical constructs.

• Self-consistency as an implication. Classic multi-rater consistency says the same image must get the same mask, d_I(I_i,I_j)=0 \Rightarrow d_M(M_i,M_j)\le\epsilon_M. The single-rater version relaxes “the same image” to “a near-identical patch”: d_I\big(I^{(p)}_i, I^{(p)}_j\big)\le\epsilon_I \;\Longrightarrow\; d_M\big(M^{(p)}_i, M^{(p)}_j\big)\le\epsilon_M . A flagged pair is one where the left side holds and the right side fails. Here d_I is the mean squared error between image patches and d_M = 1-\mathrm{IoU} between mask patches. The threshold \epsilon_I = 10^{-3} on intensities in [0,1] is a PSNR of 10\log_{10}(1/10^{-3}) = 30 dB, a difference the eye cannot see.
• Bishop (rotation-minimising) frames. To cut a cross-section you need two directions perpendicular to the centreline at every point. The Frenet–Serret frame takes them from the direction the curve bends; on a nearly straight segment that direction is undefined, and torsion makes it spin, so the same vessel gives patches rotated by arbitrary amounts. The Bishop frame carries the normals along by parallel transport. Its equations, T' = k_1N_1 + k_2N_2, N_1' = -k_1T, N_2' = -k_2T, have no N_1\leftrightarrow N_2 term, so the normals turn only as much as the tangent forces them to. More patches then line up pixel for pixel, which matters because d_I is a raw MSE with no rotation search.
• A baseline for how much correct masks disagree. Even correct masks disagree a little, and more so when the images differ more. So the neighbour pairs are binned by d_I into K=100 equal-width bins over [0,\epsilon_I), and each bin gets the mean and standard deviation of d_M: \mu_k\approx\mathbb E[d_M\mid d_I\in B_k] and \sigma_k\approx\sqrt{\operatorname{Var}[d_M\mid d_I\in B_k]}. This is a binned regression of d_M on d_I whose spread is allowed to change with d_I.
• Pair residual, a within-bin z-score. r_{ij} = \frac{\mu_k - d_M\big(M^{(p)}_i, M^{(p)}_j\big)}{\sigma_k + \varepsilon},\qquad d_I\big(I^{(p)}_i, I^{(p)}_j\big)\in B_k . The sign is flipped so that a strongly negative r_{ij} means “these masks disagree far more than masks of equally similar images usually do”.
• Patch score by median: assigning blame. r_{ij} is symmetric in i and j, so on its own it cannot say which mask is wrong. The patch score asks all of i’s near-twins, \mathcal N_i=\{\ell : d_I(I_i,I_\ell)<\epsilon_I\}, and takes the median: R_i = \operatorname{median}\{r_{ij} : j\in\mathcal N_i\}. If patch i is clean, most of its twins are clean too and the residuals scatter around 0. If i’s mask is wrong, it disagrees with nearly every twin and R_i\ll 0. The median tolerates up to half of the twins being noisy themselves (its breakdown point is 50%).
• From patches to a loss weight. Each voxel takes the score of its nearest centreline node (a Voronoi assignment), q(v)=\sigma(R(v)), and Q(v) = 1-\big(1-q(v)\big)\,w(v),\qquad \mathcal L_{qw} = \frac{\sum_v Q(v)\,\ell\big(\hat y(v),y(v)\big)}{\sum_v Q(v)}, where w(v) decays away from the vessel so background voxels keep weight 1.
• Scale. About 3\times10^6 patches, so about 10^{12} candidate pairs; a FAISS top-k search over flattened patches makes this a one-off 6-hour precomputation.

Interactive: T-AM-063 widget.

What to watch. \sigma(0)=0.5, so an ordinary clean patch (R\approx0) gets Q = 1-0.5\,w(v). Right next to the vessel, clean voxels weigh half as much as background, and only better-than-average agreement (R>0) restores full weight. The map therefore also re-balances vessel against background, not only noisy against clean. The baseline (\mu_k,\sigma_k) is estimated from all pairs, noisy ones included, so noise inflates its own \sigma_k and hides its outliers; the widget shows a median/MAD baseline scoring the same noisy patch about three times more negative. Pairs are not independent, since a patch sits in many pairs. The principle detects inconsistency: an error that is made the same way on every similar patch leaves no disagreement and is invisible. Recall depends on retrieval: rotated or window-shifted twins are missed (the paper’s own failure cases), and rare geometry such as bifurcations has few twins and so no score.

T-AM-171 · Learning to Distort: Weakly-Supervised Image Quality Transfer for Prostate DWI Correction

Yucheng Tang, Wen Yan, Alexander Ng, … Shaheer Ullah Saeed, Veeru Kasivisvanathan, Yipeng Hu — arXiv:2606.18869 · MICCAI page

What is graded. Image-level labels (distorted or not) are turned into supervision for a generator; the paper’s goal is correction.

Core idea. There are no paired distorted and undistorted DWI scans, so a corrector cannot be trained directly. The paper turns the problem around. Learning to add susceptibility distortion is safer than learning to remove it, because a hallucinated error in a degraded image does less harm than a hallucinated lesion in a “corrected” one. A flow-matching model is trained to carry undistorted volumes towards distorted ones. A quality classifier’s feature space, with one prototype per class, pushes the late part of each generated trajectory towards the “distorted” prototype. The synthetic pairs then train an ordinary supervised corrector.

Mathematical constructs.

• A quality feature space. A 3D ResNet-18 f_\phi maps a volume to z\in\mathbb R^{512}. It is trained with BCE plus a supervised contrastive loss, which pulls same-class features together on the unit sphere: \mathcal L_{con} = \frac{1}{\lvert B\rvert}\sum_{i}\frac{-1}{\lvert P(i)\rvert}\sum_{p\in P(i)}\log\frac{\exp(\cos(z_i,z_p)/\tau)}{\sum_{j\neq i}\exp(\cos(z_i,z_j)/\tau)},\qquad \tau=0.1, where P(i) holds the other samples of i’s class.
• Two quality prototypes, by margin ranking. With f_\phi frozen, vectors p_0 (undistorted) and p_1 (distorted) are learned so that every feature is at least m more cosine-similar to its own prototype than to the other: \mathcal L_{proto} = \frac1{\lvert B\rvert}\sum_i \max\big(0,\ \cos(z_i,p_{1-c_i}) - \cos(z_i,p_{c_i}) + m\big),\qquad m=0.2 .
• Flow matching with minibatch OT pairing (see optimal transport). Within each minibatch, undistorted x_i and distorted y_j are paired by an optimal transport plan. The straight path \hat y^t = (1-t)x_i + t y_j has constant velocity u = y_j - x_i, and a network learns it: \mathcal L_{FM} = \mathbb E_{(x_i,y_j),\,t\sim U(0,1)}\big\lVert \hat v_\theta(\hat y^t, t) - (y_j - x_i)\big\rVert_2^2 . Generation is Euler integration, x^k = x^{k-1} + \hat v_\theta(x^{k-1}, t_{k-1})/K, with K=20.
• Prototype guidance on the generated trajectory. During training the model is also rolled out, each late state x^k (k\ge k_0=10) is encoded, z^k=f_\phi(x^k), and pushed towards p_1 with the same hinge: \mathcal L_{PG} = \sum_{k=k_0}^{K} w_k\,\max\big(0,\ \cos(z^k,p_0) - \cos(z^k,p_1) + m\big),\qquad w_k = \Big(\frac{k-k_0}{K-k_0}\Big)^2 . The quadratic weight grows from 0 at k_0 to 1 at K, because early states are still noisy and the encoder’s reading of them is unreliable. The total loss is \mathcal L_{FM} + \lambda_{PG}\mathcal L_{PG} with \lambda_{PG}=1, and gradients are accumulated every 3 steps to fit the rollout in memory.
• Stage two. The synthetic pairs (x_i, \hat y_i = x_i^K) train a 3D U-Net corrector with \mathcal L_{corr} = \mathbb E\lVert x_i - G_\psi(\hat y_i)\rVert_1.
• Realism, measured by harm. A distortion counts as realistic if a frozen PI-RADS/Gleason classifier (ProFound) does worse on it: lower downstream accuracy and AUC are better for the generator.

Interactive: T-AM-171 widget.

What to watch. The hinge saturates too. Once \cos(z,p_1)-\cos(z,p_0)\ge m, \mathcal L_{PG} and its gradient are exactly zero, so prototype guidance stops at a fixed cosine gap of 0.2. The ablation’s explanation, that BCE stops pulling once a sample is classified while PG “pulls deeper”, holds only inside the margin band; a logistic loss decays but never reaches zero. What PG changes is the direction of the pull (towards p_1 rather than along a classifier normal) and where it stops. “Lower downstream accuracy” rewards any degradation that confuses the classifier, including blur or destroyed lesions, so it is necessary but not sufficient for realistic susceptibility geometry. The generator is trained against the same encoder that defines “distorted”, which invites encoder-specific shortcuts, as with adversarial examples. And the corrector learns to invert the generator’s distortion, not the scanner’s; the gap between the two is what the real-distortion results have to cover.

T-PM-174 · Bridging Single Distortion Artifacts and Multifactorial Clinical Quality: Few-Shot Biparametric MRI Quality Assessment via Distortion-Trained Prototypical Networks

Yucheng Tang, Alexander Ng, Wen Yan, … Veeru Kasivisvanathan, Yipeng Hu — arXiv:2606.18872 · MICCAI page

What is graded. Whether a prostate bpMRI exam is diagnostic (PI-QUAL ≥ 4), learned from DWI distortion labels.

Core idea. Low-quality exams are rare (6% of the PRIME trial has PI-QUAL < 4), and most of the rare DWI problems are distortion (87%). So the paper trains an embedding on the plentiful, comparatively objective distortion labels by episodic prototypical meta-learning. It then freezes the embedding and recomputes the two class prototypes from just five curated examples per class of the new task: distortion at a new site, or PI-QUAL itself. T2WI enters as a second branch so the network can tell real anatomy from DWI warping. FiLM plus a gradient-reversal adversary make the DWI features insensitive to the b-value.

Mathematical constructs.

• Episodes. A 2-way 5-shot 5-query episode draws a support set (5 labelled examples per class) and a query set, and the network is trained so that prototypes built from the support classify the queries. Training on many tiny tasks rehearses deployment, where a new task arrives with five examples per class.
• Masked instance normalisation. With m=\mathbb 1[x>\bar x] and \bar x the global mean intensity, voxels inside m are standardised to zero mean and unit variance and the rest are set to 0, so background air and rectal gas cannot drive the prototypes.
• FiLM conditioning on the b-value. The b-value (1400–2000 s/mm²) is one-hot encoded as b, embedded as e_b = E^\top b, and turned into a per-channel scale and shift, [\gamma(b),\beta(b)] = W_E e_b + b_E, F' = \gamma(b)\odot F + \beta(b), initialised at \gamma=1,\beta=0. In plain words: one learned affine correction of every feature channel per b-value.
• Gradient reversal for invariance. A discriminator reads F' and predicts the b-value, \hat b = \operatorname{softmax}(W_D F' + b_D), with \mathcal L_{adv} = \mathrm{CE}(b,\hat b). The gradient reversal layer is the identity going forward and multiplies the gradient by -\alpha going back. So the discriminator gets better at reading b, while everything upstream, FiLM included, is pushed to make b unreadable: a minimax game played in a single backward pass. FiLM supplies the capacity (a per-b affine correction), the adversary supplies the objective (remove what the discriminator can see).
• Cosine prototypes. Support embeddings are \ell_2-normalised and averaged per class, P_c = \frac{1}{\lvert S_c\rvert}\sum_{i\in S_c} z_i/\lVert z_i\rVert_2, and a query is scored by a softmax over scaled cosines, P(y=c\mid x) = \frac{\exp\big(\tau\cos(z,P_c)\big)}{\sum_{c'}\exp\big(\tau\cos(z,P_{c'})\big)},\qquad \tau = 1, trained with \mathcal L = \mathrm{CE} + \lambda\,\mathcal L_{adv}, \lambda=0.5.
• Adaptation is non-parametric. At test time nothing is trained. Five curated support cases per class give new prototypes, and every later scan goes to the nearer one. For PI-QUAL the two “classes” are simply PI-QUAL ≥ 4 and < 4.
• Metrics that ignore the prior. Balanced accuracy, sensitivity at 80% specificity and specificity at 80% sensitivity are sensible with 31 positives in 483.

Interactive: T-PM-174 widget.

What to watch. With \tau=1 and two classes, P(y=1\mid x)=\sigma\big(\cos(z,P_1)-\cos(z,P_0)\big). The cosine difference lies in [-2,2], so every probability lies in [\sigma(-2),\sigma(2)] = [0.119, 0.881] and the query loss never falls below -\ln 0.881 = 0.127. Every query keeps pulling on the embedding however well it is classified, which acts like strong label smoothing; cosine prototypes usually use a scale of 10–30. The transfer to PI-QUAL has a ceiling. A frozen embedding trained only on distortion places a low-PI-QUAL exam with another cause (T2 motion, noise) near the “good” prototype, so sensitivity is capped by the share of low PI-QUAL that is distortion; the widget’s ψ slider shows the cap. The support set is hand-curated for diversity, so that choice is part of the method. And a linear discriminator enforces only what a linear read-out can see, roughly equal feature means per b-value.

W-AM-041 · MedQ-Engine: A Closed-Loop Data Engine for Evolving MLLMs in Medical Image Quality Assessment

Jiyao Liu, Junzhi Ning, Wanying Qu, Lihao Liu, Chenglong Ma, Junjun He, Ningsheng Xu — arXiv:2603.19863 · MICCAI page

What is graded. Quality perception (multiple-choice questions) and quality description (free text) by a multimodal LLM, across MRI, CT, endoscopy, fundus and histopathology.

Core idea. MLLM errors on medical IQA are concentrated in a few capability × modality cells, so annotating more data uniformly is wasteful. The engine repeats evaluate → explore → evolve. It finds the dev-set questions the model gets persistently wrong, clusters them into failure prototypes, retrieves look-alike images from a million-image pool, annotates them cheaply (GPT-4o drafts, human review only where needed), fine-tunes, and repeats. Most of the mathematics is in the selection and routing rules.

Mathematical constructs.

• The target problem, stated and set aside. Choose D_{train}\subset U with \lvert D_{train}\rvert\le B to maximise test performance of the fine-tuned model. This subset-selection problem is combinatorial, and the loop approximates it greedily, one round at a time.
• Persistent failures. Each dev sample is answered R=5 times (sampling makes answers vary) and enters the failure pool when its error rate exceeds \gamma=0.6, that is, when at least 4 of 5 answers are wrong. If a sample’s per-answer error probability is p, the chance of being flagged is a binomial tail, \Pr(\text{flag}\mid p) = \Pr\big(\mathrm{Bin}(5,p)\ge 4\big) = 5p^4(1-p) + p^5, which is 0.19 at p=0.5 and 0.74 at p=0.8. A coin-flip question is rarely flagged; a consistently wrong one usually is.
• Failure prototypes. Failures are embedded (image plus question/answer features) and clustered agglomeratively. The number of clusters is chosen by the silhouette s(i) = \frac{b(i)-a(i)}{\max\{a(i),b(i)\}}, where a is the mean distance to the own cluster and b to the nearest other cluster. The centroids p_j are the prototypes.
• Where the model is weak. For each capability dimension k, e_k = \frac{\lvert\{b\in\mathcal B : c_k(b)=1\}\rvert}{\lvert\{s\in D_{dev}: c_k(s)=1\}\rvert}, the dev error rate restricted to questions that test k.
• Retrieval and budget. Candidates are pool images whose BiomedCLIP embedding is close to a prototype’s visual part, \mathcal N(p_j) = \{x : \cos(p_j^{vis}, f_{enc}(x)) > 0.75\}. The budget is split with weights w_k\propto e_k^{\alpha}: \alpha=0 is uniform, \alpha=1 proportional to error, and large \alpha spends everything on the worst dimension.
• Routing by self-confidence. The model’s confidence in its own answer \hat y^{self} is the per-token average negative log-likelihood, H^{traj}(x,q) = -\frac{1}{\lvert\hat y^{self}\rvert}\sum_{l}\log M_{\theta}\big(\hat y^{self}_l\mid x,q,\hat y^{self}_{<l}\big), the log-perplexity of its own answer (a one-sample estimate of per-token entropy). With the agreement \delta = R(\hat y^{self},\hat y^{GPT}) there are three routes: if H\ge\tau_H, use GPT-4o’s label; if H<\tau_H but \delta<\tau_{ann}, send the sample to an expert; otherwise keep the model’s own label.
• Evolve. Deduplicate by perceptual hash, drop near-duplicate texts by TF-IDF, and fine-tune with the usual token-level NLL, \mathcal L_{SFT} = -\mathbb E\sum_i\log p_\theta(y_i\mid y_{<i},x,q).

Interactive: W-AM-041 widget.

What to watch. Two of the three routes skip humans. Route (a) adopts GPT-4o’s label exactly where the student is unsure, and GPT-4o scores only 64.8% on the perception benchmark. Route (c) passes errors that student and oracle share, and sharing grows as the student is trained on the oracle’s labels; in the widget, raising the shared-error share lowers the review rate while the silently kept wrong labels rise. The reported 18% review rate follows from the thresholds, and \tau_H, \tau_{ann}, \alpha and the agreement function R are not reported. A low H^{traj} means fluent, not correct, so confidently wrong answers go to route (c). And retrieval uses the visual half of a centroid built from joint features; a centroid need not look like any real image.

W-AM-042 · Large-Small Model Collaboration for Zero-Shot Surgical Phase Recognition

Yiyi Zhang, Ying Zheng, Wenxin Fan, … Zheng Li, Pheng-Ann Heng — arXiv:2608.22879 · MICCAI page

What is graded. The quality of pseudo-labels, not of images: a “dynamic quality control” step decides which frame labels to trust, which is why the paper carries the IQA tag.

Core idea. A surgical vision–language model labels frames zero-shot but has no sense of time, so its labels flicker. A small temporal network (MS-TCN) models time well but needs labels. LaST lets the large model label, and two small models learn from the labels they find trustworthy, cross-teaching each other, with a smoothness loss and a class-balance prior. Their ensembled output replaces the labels, and the loop runs three times. Adaptation and evaluation both use the unlabelled test videos, so the setting is transductive.

Mathematical constructs.

• Zero-shot labels. \hat y_t^{(0)} = \arg\max_{c}\ \mathrm{Sim}\big(M_L(x_t), \text{text}(P_c)\big): the phase whose text prompt is closest to the frame embedding.
• Small-loss selection with a per-class two-component GMM. Networks fit clean labels before they memorise wrong ones, so early in training a wrong label tends to have a large loss. Within each pseudo-class c, the per-frame losses are modelled as f_c(\ell) = \sum_{k=1}^{2}\pi_k\,\mathcal N(\ell\mid\mu_k,\sigma_k^2), fitted by EM (see EM and Gaussian mixtures). The component with the smaller mean is “clean”, and each frame gets the responsibility w_t = \Pr(\text{clean}\mid\ell_t) = \pi_1\mathcal N(\ell_t\mid\mu_1,\sigma_1^2)/f_c(\ell_t). Frames with w_t>\tau=0.9 form the labelled set X_L. Fitting per class stops a rare phase, whose frames all have larger losses, from being declared noise wholesale.
• Cross-learning (co-teaching, DivideMix style). Model \theta_1 trains on the clean set chosen by \theta_2 and vice versa, so one network’s confident mistakes are not fed straight back into itself.
• Three losses per model. \mathcal L^{\theta_i}_{sparse} = -\frac{1}{\lvert X_L^{\theta_j}\rvert}\sum_{x_t\in X_L^{\theta_j}}\log p_{\theta_i}(\hat y_t\mid x_t),\qquad \mathcal L^{\theta_i}_{temp} = \frac{1}{T-1}\sum_{t=2}^{T}\big\lVert\log p_{\theta_i}(\cdot\mid x_t) - \mathrm{sg}[\log p_{\theta_i}(\cdot\mid x_{t-1})]\big\rVert_2^2, \mathcal L^{\theta_i}_{reg} = D_{KL}(\phi\,\Vert\,\bar p_{\theta_i}) = \sum_{c}\frac1C\log\frac{1/C}{\bar p_{\theta_i,c}},\qquad \bar p_{\theta_i} = \frac{1}{\lvert V\rvert}\sum_{x\in V}p_{\theta_i}(\cdot\mid x). The temporal term pulls each frame towards its predecessor, and the stop-gradient makes the pull one-way, forwards in time. The KL term compares the video-average prediction with a uniform prior. It grows without bound if any phase’s average probability goes to 0, which is what prevents collapse onto one phase. The unlabelled set X_U enters no loss, hence “sparse”.
• Cycle replay. After a cycle, the ensemble of \theta_1 and \theta_2 relabels the video, \hat Y^{(r)} replaces \hat Y^{(r-1)}, and the GMM split is redone on the better labels, for r=3 cycles.

Interactive: W-AM-042 widget.

What to watch. The class-balance prior is uniform, but surgical phases are not: in a cholecystectomy two dissection phases take most of the time. A toy calculation makes the pull visible. If the only other pressure were a cross-entropy to the true phase shares q, minimising -\sum_c q_c\log\bar p_c + \lambda D_{KL}(u\Vert\bar p) gives \bar p = (q+\lambda u)/(1+\lambda), so at \lambda = 1 the predicted shares sit halfway between truth and uniform. The real losses act per frame, so the pull is weaker, but it points the same way: short phases get lengthened. Loss-based selection finds labels that are hard to fit, not labels that are wrong. A confusion the foundation model makes systematically (two phases that look alike) is easy for the small model to fit, has a low loss, and is selected as clean; the widget shows the kept set’s clean ratio falling as such labels increase. And a per-class GMM on a rare class has few points to fit.

W-PM-008 · GeoCAN: Nonlinear Causal-Geometric Learning for Echocardiography Quality Assessment

Yiran Li, Kai Zheng, Shuo Li — MICCAI page (no preprint listed in the paper list)

What is graded. An echocardiography image’s quality grade (11 levels on CACTUS, 5 on the other datasets), jointly with its view.

Core idea. Quality should be read from the relations between visual factors (chambers, walls, valves), not from appearance that merely co-occurs with good images. GeoCAN builds a directed factor-by-factor matrix from how much of each factor map’s activation is covered by another’s. It turns the asymmetric part of that matrix into a per-factor gain and adds a hyperbolic ranking loss on multi-hop relations. The prediction head is a Kolmogorov–Arnold-style spline network. “Causal” here means directed co-activation; no intervention is modelled.

Mathematical constructs.

• A tunable pooling operator (generalised Lehmer mean). For a spatial map F_k with S positions, \mathrm{GLM}_{\alpha,\beta}(F_k) = \frac{1}{\ln\alpha}\ln\frac{\sum_s\alpha^{(\beta+1)F_k(s)}}{\sum_s\alpha^{\beta F_k(s)}} . Two readings help. First, with x_s=\alpha^{F_k(s)} it is \log_\alpha of the Lehmer mean L_p(x)=\sum_s x_s^p/\sum_s x_s^{p-1} at p=\beta+1, and Lehmer means run from the minimum (p\to-\infty) through the arithmetic mean (p=1) to the maximum (p\to\infty). Second, with \kappa=\ln\alpha it equals \frac1\kappa\ln\sum_s\pi_s e^{\kappa F_k(s)} with weights \pi=\operatorname{softmax}(\beta\kappa F_k). So \beta decides which positions the pool listens to, \kappa how max-like the average over them is, and the result always lies between \min_s F_k and \max_s F_k. It tends to the max as \kappa\beta\to+\infty, to the min as \kappa\beta\to-\infty, and to the mean as \kappa\to0.
• A “conditional” matrix. By analogy with P(A\mid B) = P(A\cap B)/P(B), with GLM playing the role of a soft measure, C_{ij} = \frac{\mathrm{GLM}_{\alpha,\beta}\big(F^{(i)}\odot F^{(j)}\big)}{\mathrm{GLM}_{\alpha,\beta}\big(F^{(j)}\big)+\varepsilon} . Near mean pooling this is “how much of j’s activation also lies under i”. If a large factor i covers a small factor j, C_{ij} is large and C_{ji} small: the matrix is asymmetric.
• A tournament from the asymmetric part. With A = C - C^\top (antisymmetric) and P=\sigma(A/\tau), every pair satisfies P_{ij}+P_{ji}=1, so P_{ij} reads as the probability that factor i “beats” factor j in a round-robin tournament. The per-factor score is g = \mathrm{Norm}\Big(\sigma\big(\tfrac{A}{\tau}\big)\big(1-\sigma(\tfrac{A}{\tau})\big)^{\!\top}\mathbf 1\Big) . Because 1-\sigma(a)=\sigma(-a) and A^\top=-A, the second factor equals \sigma(A/\tau) itself, so g=\mathrm{Norm}(P^2\mathbf 1): each factor’s wins, weighted by how many wins its victims have (a Kendall–Wei-style second-order tournament score). The features are then rescaled factor by factor, X_{cgl} = X\odot(1+r\tanh g), a gain between 1-r and 1+r.
• Hyperbolic embedding and multi-hop ranking. Factor embeddings are mapped into the Poincaré ball by h_i=\exp_0^c(z_i) = \tanh(\sqrt c\lVert z_i\rVert)\,z_i/(\sqrt c\lVert z_i\rVert). Multi-hop relations come from the SVD C=USV^\top. Even powers give (CC^\top)^n = US^{2n}U^\top, and a \lambda-weighted series of them gives \mathrm{hop} = U\operatorname{diag}(S')U^\top. A softplus (BPR-style) ranking loss asks dominant pairs to score above suppressed ones, \mathcal L_{geo}=\sum\log\big(1+\exp(\mathrm{hop}(i,j^-)-\mathrm{hop}(i,j^+))\big).
• Spline aggregation (KAN-style). Each token channel passes through learnable univariate spline functions, \mathcal F(t)=\sum_{j=1}^{D}\sum_{m=1}^{M}w_{jm}\phi_{jm}(t_j), which are added residually, f' = f+\beta\sum_i\psi_i\mathcal F(t_i), and the score is \hat y = \sum_m u_m\psi_m(f'). A Kolmogorov–Arnold network replaces the usual MLP head.

Interactive: W-PM-008 widget.

What to watch. Nothing here is causal in the do-calculus sense. C_{ij} is a co-activation ratio, so it cannot tell a cause from a confounder, which is the problem the paper sets out to solve; the widget also shows the “direction” changing with the pooling knob alone. The (1-\sigma)^\top factor in Eq. (3) adds nothing beyond \sigma for an antisymmetric argument. The multi-hop matrix U\operatorname{diag}(S')U^\top is symmetric, so the direction that C-C^\top was built to keep is discarded by the ranking loss. How j^+ and j^- are chosen is not stated, and as printed Eq. (5) uses \mathrm{hop} rather than hyperbolic distances between the h_i, so the link between the hyperbolic embedding and the loss is not written down. With a learnable \alpha, \ln\alpha\to0 is a removable but numerically delicate singularity.

W-PM-121 · SlideGuard: WSI Artifact Detection Benchmark

Gabriela Kaczmarek, Zuzanna Krawczyk-Borysiak, Mateusz Miller, … Zaneta Swiderska-Chadaj — MICCAI page · oral · dataset · benchmark

What is graded. Seven existing whole-slide-image QC tools, on 53 H&E slides (4 sources, 2 species, 6 tissues) with 2,116 pixel-level annotations of 11 artefact types.

Core idea. A benchmark, not a method. Every tool runs with its default settings. Pixel-mask tools are scored by Dice inside the tissue region, and patch-classification tools against a ground truth discretised to their own grid. The headline finding, that no tool exceeds 0.6 artefact Dice while artefact-free Dice is high, is largely a property of the metric under heavy class imbalance, which is why the metric is worth unpacking.

Mathematical constructs.

• Evaluation region. Tissue is the pixels darker than 200 (8-bit), cleaned by morphological opening and closing and checked by hand. Every metric is computed inside it, so tools are not scored on their own tissue detection.
• Two Dice scores from one confusion matrix. Artefact Dice is \frac{2TP}{2TP+FP+FN}, and artefact-free Dice \frac{2TN}{2TN+FP+FN} is the Dice of the inverted masks. Write the artefact share of tissue as \pi, and a detector’s sensitivity and false-positive rate as s and f. Per unit of tissue, \mathrm{Dice}_{art} = \frac{2\pi s}{2\pi s + (1-\pi)f + \pi(1-s)},\qquad \mathrm{Dice}_{free} = \frac{2(1-\pi)(1-f)}{2(1-\pi)(1-f) + (1-\pi)f + \pi(1-s)} . At \pi = 2\%, a detector with 80% sensitivity and 98% specificity scores \mathrm{Dice}_{art}=0.58 and \mathrm{Dice}_{free}=0.99. Its false positives are only 2% of the clean tissue, but that is as much area as the artefacts themselves. The same detector at \pi=20\% scores 0.85. The “detection asymmetry” the paper reports is what this formula predicts.
• Patch discretisation. For patch-level tools a patch counts as artefact if annotations cover more than 25% of it, predictions are thresholded at 0.5, and Dice becomes the patch F1. Thin artefacts (knife marks, scratches, narrow folds) rarely fill a quarter of a patch, so they drop out of the patch ground truth. That is why pixel- and patch-mode scores cannot be compared.
• Label-agnostic recall. Each tool has its own artefact taxonomy, so per-type precision is undefined. The paper reports, per annotated type, the share of its pixels flagged as any artefact.

Interactive: W-PM-121 widget.

What to watch. Recall without precision rewards over-flagging: a tool that marks everything scores recall 1 on every type (PathProfiler reaches 1.00 on air bubbles). Per-slide Dice is unstable when a slide’s artefact area is tiny, which is why the near-artefact-free “HQ” slides switch to artefact-free Dice. All annotations come from one histotechnologist, so there is no inter-rater ceiling to compare 0.6 against, and for soft-edged artefacts the boundary itself is uncertain. The intensity-200 tissue mask can clip pale artefacts such as bubbles or faint folds. Running every tool at its defaults is impartial but also measures configuration fit; as the authors point out, a nominal “20x” means anything from 0.25 to 0.50 µm per pixel across scanners.

W-PM-123 · Revisiting Integration of Image and Metadata for DICOM Series Classification: Cross-Attention and Dictionary Learning

Tuan Truong, Melanie Dohmen, Sara Lorio, Matthias Lenga — arXiv:2602.23833 · MICCAI page

What is graded. Which MRI series a DICOM series is (T1, T2, DWI, ADC, Dixon, plane, contrast phase). This is a prerequisite for automated QC and protocol harmonisation rather than image quality itself.

Core idea. The DICOM header is informative but patchy: tags are vendor-specific, edited by hand, or missing. Instead of imputing missing tags, each slice’s metadata is treated as a set of the (tag, value) pairs that are actually present. Each pair is embedded as a learnable per-tag vector modulated by the value, and the set is averaged. Image tokens (a 2.5D encoder over 10 equidistant slices) and metadata tokens then attend to each other in both directions before a learned pooling.

Mathematical constructs.

• A series as a set of slices. From N slices, S=10 equidistant ones are kept, giving images x\in\mathbb R^{S\times H\times W} and metadata y\in\mathbb R^{S\times F} with NaN for missing entries. Slice tokens attend to each other across slices.
• Sparse metadata encoder: a set encoder with FiLM. For slice s with observed tags O_s, each tag f has a learnable embedding e_f\in\mathbb R^d, and a small value network reads the value and the tag and returns a scale and a shift: (\alpha_{s,f},\beta_{s,f}) = g_\theta\big([v_{s,f}, e_f]\big),\qquad \tilde e_{s,f} = e_f\odot(1+\alpha_{s,f})+\beta_{s,f},\qquad m_s = \mathrm{MLP}\Big(\frac{1}{\lvert O_s\rvert}\sum_{f\in O_s}\tilde e_{s,f}\Big). A missing tag contributes nothing, rather than a zero, and the mean works for any number of observed tags in any order (a Deep Sets-style encoder). The tag identity says what the number is (echo time, flip angle); FiLM lets the same number mean different things for different tags.
• Bi-directional cross-attention. After projection to a shared width, V' = \mathrm{MHA}(Q=\tilde V, K=\tilde M, V=\tilde M) and M' = \mathrm{MHA}(Q=\tilde M, K=\tilde V, V=\tilde V), each followed by a residual connection, a feed-forward block and LayerNorm, and then F=\mathrm{GELU}(\mathrm{LN}([V'',M'']W_f)).
• Learned slice pooling. z=\sum_{s=1}^S w(F)_s F_s with w:\mathbb R^{S\times d_o}\to[0,1]^S. These are gates rather than softmax weights, so they need not sum to one.

Interactive: W-PM-123 widget.

What to watch. Mean pooling forgets how many tags were present. When the pattern of missing tags is itself informative (vendor, sequence family), only the identities of the present tags carry that signal; a sum or an explicit count would keep it. The preprocessing section also appends a binary missingness indicator per categorical tag, which re-introduces missingness as an observed value, and which encoders see it is not spelled out. Most header tags are constant within a series, so the S metadata tokens are near-copies, and cross-attention over them adds little beyond the slice-varying tags. The pooling gates are unnormalised, so the series embedding’s scale grows with S. Weighted F1 is dominated by the frequent classes, and the out-of-domain drops (Dixon opposed-phase 74%, portal venous 65%) look as much like label-definition shift as like covariate shift.

W-PM-139 · FunPiQ: A New Benchmark for Pixel-Level Quality Assessment in Fundus Images

Pengwei Wang, José Morano, Virginia Mares, Hrvoje Bogunović — arXiv:2606.25915 · MICCAI page · code and data

What is graded. Every pixel of a colour fundus photograph, as good, usable or bad, according to whether the anatomy under it is visible.

Core idea. Image-level quality labels disagree across datasets because “gradable” depends on the downstream task. Pixel-level visibility is task-agnostic, and a task-specific rule (“is the optic disc in a good region?”) can be layered on top. The paper releases 300 pixel-annotated images from EyeQ, BRSET and mBRSET, two thirds from Brazil and some from mobile cameras, as an evaluation-only benchmark. It also proposes EFIQA-CP, which trains a small convolutional adapter on frozen DINOv3 features from pseudo-labels that say “no vessels are visible here”. Only confidently bad pixels are trusted and everything else is treated as unlabelled, so it trains with non-negative positive-unlabelled (nnPU) learning.

Mathematical constructs.

• Positive-unlabelled learning in plain words. We have some pixels known to be bad (positives) and many we know nothing about (unlabelled, a mix of bad and good). Treating every unlabelled pixel as good would teach the model that the hidden bad pixels are good. PU learning instead uses the class prior \pi (the share of bad pixels, set to 0.05) to subtract the bad pixels’ expected contribution from the unlabelled set.
• The unbiased PU risk. The unlabelled density is the mixture p(x)=\pi p_+(x)+(1-\pi)p_-(x), so the loss on true negatives, which we cannot observe, can be written with quantities we can: (1-\pi)\,\mathbb E_-[\ell(g(x),-1)] = \mathbb E_u[\ell(g(x),-1)] - \pi\,\mathbb E_+[\ell(g(x),-1)]. The classification risk becomes \hat R_{pu}(g) = \pi\hat R_p^+ + \hat R_u^- - \pi\hat R_p^-, where \hat R_p^+ and \hat R_p^- are the losses on labelled positives scored as positive and as negative, and \hat R_u^- is the loss on unlabelled pixels scored as negative.
• Why it must be clamped. The last two terms estimate a loss, which cannot be negative, but they are a difference of two sample averages. A flexible network can drive the difference below zero by memorising the labelled positives, and minimising \hat R_{pu} then rewards exactly that overfitting. nnPU clamps the estimate (the paper’s Eq. 1, with a softplus surrogate loss), \mathcal L = \pi\hat R_p^+ + \max\big(0,\ \hat R_u^- - \pi\hat R_p^-\big), and the training rule of Kiryo et al. (2017) takes a step that raises the negative-class term whenever it dips below zero.
• Adapter with context. Five ConvNeXt blocks with 7 × 7 depthwise convolutions and MLP ratio 1 (instead of 4) give a large receptive field with little capacity. The fovea, which has few vessels by anatomy, can then be told apart from a region where vessels are hidden.
• Ordinal agreement: quadratic weighted kappa. With observed confusion counts O_{ij}, the table E_{ij} expected from the two marginals alone, and weights w_{ij}=(i-j)^2/(C-1)^2, \kappa_w = 1-\frac{\sum_{ij}w_{ij}O_{ij}}{\sum_{ij}w_{ij}E_{ij}} . Calling a bad pixel good costs four times as much as calling it usable, and chance agreement scores 0. A binary “reject” task (bad against the rest) adds Dice, AUROC, AUPRC and sensitivity at 95% specificity.

Interactive: W-PM-139 widget.

What to watch. PU theory assumes the labelled positives are a random sample of all positives (“selected completely at random”). Here they are the pixels the pseudo-labeller was most certain about, the extreme tail. The classifier then learns what very bad looks like, and the \pi\hat R_p^- correction subtracts the wrong amount; in the widget even the best possible classifier gets a negative uPU value under strict selection. The prior \pi=0.05 is set, not estimated, and a wrong \pi moves the decision threshold. Each method’s metrics are reported at its own best threshold, which inflates absolute numbers for everyone. Labels come from expert readers under one ophthalmologist’s supervision, with multi-reader agreement left for future work.

W-PM-194 · ICHOR: A Robust Representation Learning Approach for ASL CBF Maps with Self-Supervised Masked Autoencoders

Xavier Beltran-Urbano, Yiran Li, Xinglin Zeng, … John A. Detre, Sudipto Dolui — arXiv:2603.05247 · MICCAI page · spotlight

What is graded. One of the four downstream tasks regresses a quality score in [0,1] for ASL cerebral-blood-flow (CBF) maps; the paper itself is a pretraining method.

Core idea. No pretrained backbone exists for perfusion MRI, and models pretrained on structural MRI learn edges and tissue boundaries that CBF maps lack. ICHOR pretrains a 3D ViT-B as a masked autoencoder on 11,405 CBF maps from 14 studies, then adapts it to each small labelled task with LoRA.

Mathematical constructs.

• Patches and masking. A 96^3 map with patch size P=12 gives N=(96/12)^3=512 patches of 12^3=1728 voxels. A random half (\rho=0.5) is hidden, and the encoder sees only the 256 visible tokens, which is what makes MAE pretraining cheap.
• Asymmetric encoder and decoder. The encoder is a ViT-B (12 blocks, width 768, sinusoidal positions). A light decoder (4 blocks, width 384) receives all 512 positions: the encoded visible tokens, plus one shared learnable mask token at every hidden position. A linear head predicts each patch’s 1728 voxels.
• Loss only where the answer was hidden. \mathcal L_{MSE} = \frac{1}{\lvert\mathcal M\rvert}\sum_{i\in\mathcal M}\lVert\hat p_i - p_i\rVert_2^2 . Scoring the visible patches would reward copying the input.
• LoRA adaptation. Each attention projection W_0\in\mathbb R^{768\times768} (query, key, value, output) becomes W_0 + \frac{\alpha}{r}BA with B\in\mathbb R^{768\times r}, A\in\mathbb R^{r\times768}, r=8 and \alpha=16, and only A and B are trained. That is 12\times4\times2\times768\times8 = 589{,}824 parameters, about 0.7% of the roughly 86M in the encoder. Features are global-average-pooled into a LayerNorm and a linear head.
• The masking-ratio trade-off. With too little masking a hidden patch is predictable from its neighbours; with too much there is not enough context. On the amyloid task \rho=0.5 beat 0.25 and 0.75 (AUC 78.9 against 76.7 and 71.9), lower than the 0.75 usual for natural images.

Interactive: W-PM-194 widget.

What to watch. CBF maps are smooth at the 2 mm scale, so part of the pixel-MSE pretext task can be solved by interpolating from visible neighbours, without learning anything about perfusion patterns. In the widget’s toy, interpolation alone explains about 80% of the hidden-patch variance at \rho=0.5. The ρ comparison uses one task of 147 subjects with five folds, where a 2-point AUC gap is within noise. The pretraining set was filtered by an automated QC tool, so the encoder has mostly seen acceptable maps, which may be why its advantage is smallest on the quality task. Cohorts of 35 and 63 subjects (AUC 100% on AD against bvFTD) make the classification numbers fragile. Adapting the structural-MRI baselines with LoRA rather than full fine-tuning partly confounds the comparison, as the authors note.

W-PM-212 · AC-MIL: Weakly-Supervised Atrial LGE-MRI Quality Assessment via Adversarial Concept Disentanglement

K M Arefeen Sultan, Kaysen Hansen, Benjamin Orkild, … Ed DiBella, Shireen Elhabian — arXiv:2604.10303 · MICCAI page

What is graded. A 4-point quality grade for atrial LGE-MRI volumes, decomposed into three clinical concepts (sharpness, myocardium nulling, aorta and valve enhancement) plus a residual, from volume-level labels only.

Core idea. Multiple-instance learning can grade a volume from a single label, but it squeezes all the evidence into one opaque vector. AC-MIL splits that vector into four concept slots, each with its own attention over patches. Three slots are trained only by their own concept grades, and a stop-gradient keeps the task loss out of them. The fourth, residual slot carries whatever else the task needs. Adversaries stop it from re-learning the three concepts, and a penalty stops the localised concepts’ attention maps from overlapping.

Mathematical constructs.

• Nested bags. A volume is a bag of the M slices that contain the left atrium. Training samples N=8 slices, and each slice is a bag of K=80 random 64 × 64 patches. Only the volume carries labels.
• Concept slots with gated attention. A shared ResNet gives h_k, and four projection heads give h_k^{(c)} for c\in\{sh,nu,ao,un\}. Each slot pools the patches with its own gated attention (Ilse et al.), a_k^{(c)} = w_c^\top\big(\tanh(V_ch_k^{(c)})\odot\sigma(U_ch_k^{(c)})\big),\qquad \alpha^{(c)} = \operatorname{softmax}_k\big(a^{(c)}\big),\qquad Z^{(c)} = \sum_k\alpha_k^{(c)}h_k^{(c)} .
• Asymmetric stop-gradient. Slices are fused as V_m=[\mathrm{sg}(Z_{sh}),\mathrm{sg}(Z_{nu}),\mathrm{sg}(Z_{ao}),Z_{un}], pooled over slices by a second attention, V_{vol} = \sum_m\beta_mV_m, and graded. The task gradient reaches only the residual slot, so the three concept slots cannot quietly absorb task information (“concept leakage”). It still reaches the shared ResNet through that slot.
• CORN ordinal loss. For K ordered grades, K-1 binary heads estimate conditional probabilities f_k(x) = \Pr(y>k\mid y>k-1), and head k is trained only on samples with y>k-1. The cumulative probabilities are products, \Pr(y>k)=\prod_{j\le k}f_j(x), so they can only decrease as k grows (rank consistency by construction). The predicted grade is 1+\sum_k\mathbb 1[\Pr(y>k)>0.5]. CORN is used for the task and for each concept.
• Adversarial erasure. Adversaries D_c try to predict each concept grade from Z_{un}, \mathcal L_{adv} = \sum_c\mathcal L_{CORN}(D_c(Z_{un}),y_c), through a gradient reversal layer: the adversaries get better, and the residual slot learns to defeat them.
• Spatial attention diversity. For the three localised slots (sharpness is global and excluded), \mathcal L_{SAD} = \sum_{c_i\neq c_j\in\{nu,ao,un\}}\frac{\langle\alpha^{(c_i)},\alpha^{(c_j)}\rangle}{\lVert\alpha^{(c_i)}\rVert_2\lVert\alpha^{(c_j)}\rVert_2} .
• Total. \mathcal L = \mathcal L_{task} + 1.0\,\mathcal L_{CBM} + 0.5\,\mathcal L_{adv} + 0.1\,\mathcal L_{SAD}, evaluated by QWK and class-averaged MAE.

Interactive: W-PM-212 widget.

What to watch. Attention weights are non-negative, so the cosine between two attention maps is at least 0 and reaches 0 only when their supports are disjoint. Two near-uniform maps have cosine near 1 whatever they attend to, so \mathcal L_{SAD} pushes towards peaky, disjoint attention, not just different attention. Training uses the cosine, while the evaluation reports Jensen–Shannon divergence. An adversary at chance shows that that adversary cannot decode the concepts from Z_{un}, not that Z_{un} is “statistically independent” of them. Adding the diversity term to the concept-plus-adversary model lowers QWK from 0.68 to 0.66, so the more interpretable variant is not the most accurate one. Concept grades have an inter-rater ICC of 0.62–0.72, which bounds how well the concept slots can be supervised.

Threads across the papers

• Prototypes are the default way to use few labels. T-AM-171 learns two prototypes as targets for a generator, T-PM-174 recomputes them from five cases per class, and MedQ-Engine uses cluster centroids as retrieval anchors. All three score by cosine similarity. In two of them a margin or a scale decides whether a well-placed sample ever stops being pulled: T-AM-171’s hinge stops exactly at the margin, and T-PM-174’s τ = 1 never lets the pull stop. That detail is worth checking whenever a paper argues about “pulling deeper” or “saturating”.
• Weak labels need a model of how they are wrong. FunPiQ’s nnPU assumes random selection of the labelled positives, LaST’s GMM assumes wrong labels are hard to fit, and T-AM-063’s median assumes most twins are clean. Each fails in a specific, predictable way (extreme-tail labels, systematic confusions, majority-noisy neighbourhoods), and each paper’s own setting comes close to its failure case.
• Gradient reversal appears twice with the same overclaim risk. In T-PM-174 and AC-MIL, a discriminator at chance is read as invariance or independence. It shows only that a model of that capacity cannot read the nuisance.
• The metric shapes the headline. SlideGuard’s sub-0.6 ceiling is largely class imbalance in Dice; FunPiQ reports each method at its best threshold; T-AM-171 measures realism by how much a classifier is harmed. Reading the metric’s formula first changes how the results read.

Takeaways

• The strongest ideas here are the statistical ones: T-AM-063’s model-free consistency test with explicit evidence, FunPiQ’s move to a positive-unlabelled risk, and T-PM-174’s honest use of a cheap proxy label (distortion) for an expensive one (PI-QUAL).
• The weakest claims are about causality and independence (GeoCAN, AC-MIL): the mathematics shown measures co-activation or adversary failure, not either of those.
• For applied IQA work, the most transferable pieces are pixel-level quality maps trained from anatomy-visibility pseudo-labels with nnPU, few-shot prototypes on a proxy-trained embedding, and a failure-driven data loop, each with the caveats listed above.